A fisheye camera has a feature that a single image having a wide viewing angle of more than 180 degrees can be taken and acquired, and because of this feature, the fisheye camera is used in an industrial field such as monitoring, a map generation, and an automotive rear-view monitor.
However, it is known that a fisheye camera generates a large optical distortion in a taken image. Therefore, if the fisheye camera is used in a task such as detecting a specific object such as a pedestrian in an image, the detection is sometimes difficult.
For example, as a detection method of the object, the detection methods by the machine learning using the statistical discriminator is generally known, but those are determination methods based on the similarity of a feature such as shape pattern of a pedestrian. Thus, it is difficult to apply the statistical identification that is designed on the assumption of an image taken by an ordinary camera (i.e. an optical system that can approximate a pinhole model and a perspective projection model) to a fisheye image in which the shape of the same object largely varies depending on a position in an image due to the optical distortion.
A method of correcting the optical distortion of a fisheye image and presenting a target in easily viewable way is proposed. This method includes modeling a fisheye image as mapping to the plane of the virtual image projected on the sphere surface, cutting out and converting a part of it into a corrected image (referred to as “perspective projection correction image”) that approximates to an image taken according to the perspective projection model.
Although a straight line in a real space is restored as a straight line in the perspective projection correction image and the perspective projection correction image accurately approximates to a perspective projection camera image at a central part, the perspective projection correction image is an image in which an object is more unnaturally enlarged due to the projection distortion at a peripheral part as it is more peripheral. In the perspective projection correction image, it is not possible in principle to project an image from a viewpoint in any direction at an angle exceeding 90 degrees from an optical axis whose angle is 0 degree. Accordingly, the perspective projection correction image can be used only for a use application that requires only an image of a central part having small projection distortion.
To solve such problems, for example, PTL 1 and PTL 2 disclose a technology for correcting a distortion of a person and the like taken in a fisheye camera image.
Specifically. PTL 1 discloses an image converting device. The image converting device disclosed in PTL 1 models a fisheye image on a virtual sphere surface, projects it on a cylindrical coordinate system, generates a curved regular image, thereby correcting distortion. In the cylindrical coordinate system, a horizontal direction is set to be a tangential direction of a sphere surface and a vertical direction is set to be an axial direction of the sphere surface.
PTL 2 discloses an image presentation device using the image converting device disclosed in PTL 1. An image processing device disclosed in PTL 2 corrects, using the curved regular image disclosed in PTL 1, an image of a plurality of persons in a range of a view field extending in a horizontal direction from a central view field of the fisheye image.